Bit Extraction Technique: Methods & Applications

• Bit extraction technique is a family of methods that isolate, reconstruct, and recycle bit-level information from structured representations across multiple domains.
• It employs methodologies like shift-and-mask, canonical decomposition, and symbolic elimination to achieve unbiased and efficient data processing.
• Applications span RNG scaling, image steganography, GF arithmetic, and SIMD codecs, with performance gains from precise acceptance-rejection and bit recycling strategies.

Searching arXiv for relevant papers on “bit extraction technique” across domains. arXiv search query: bit extraction technique arXiv random number scaling bit plane extraction ANF output extraction Bit extraction technique denotes a family of procedures that isolate, reconstruct, recover, or recycle bit-level information from a structured representation. In the cited literature, the extracted object may be an unbiased sample from a random source, a bit plane or virtual coefficient in an image, a unique algebraic normal form for a circuit output bit, a continuation-flag mask in a variable-byte codec, an arbitrary-precision numeric word from decomposed storage lanes, a demodulated bit string from FSK symbols, or filtered output bits from an LFSR stream (Mennucci, 2010, Borgohain, 2012, Yu et al., 2016, Vorabbi et al., 2023, Plaisance et al., 2015, Agrawal et al., 2024, Sanchez, 2024, Nobach, 2024).

1. Formal scope and recurrent abstractions

Across these works, bit extraction is not restricted to the literal operation of masking a register. It includes direct field extraction, canonical decomposition, symbolic elimination, entropy-preserving state transformation, and inverse recovery under structural constraints. Representative forms include the binary image-plane identity

$$\mathrm{bit}_k(x,y)=\left\lfloor \frac{I(x,y)}{2^k}\right\rfloor \bmod 2 = ((I(x,y)\gg k)\$$

the weighted decomposition

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$

and the eXmY field extraction rules

$$s = (x \gg (e + m))$$

A distinct but related form is symbolic bit extraction from circuit structure, where each output bit is rewritten into a unique ANF over primary inputs by repeated substitution and simplification modulo $2$ (Borgohain, 2012, Agrawal et al., 2024, Yu et al., 2016).

Domain	Extracted object	Representative mechanism
RNG scaling	Uniform sample $Y \in \{1,\dots,M\}$	Acceptance-rejection; bit recycling reservoir
Image steganography and vision	$k$-th bit plane or weighted coefficient $a_j$	Shift-and-mask; canonical weighted decomposition
GF arithmetic verification	Unique ANF of $y_i$	Backward substitution modulo $2$
SIMD codecs	Continuation-flag mask and 7-bit payload groups	pmovmskb/vpmovmskb and PSHUFB
Arbitrary-precision quantization	Sign, exponent, mantissa fields and $N$-bit packed words	Shift/mask and lane decomposition
SDR demodulation	Hard or soft bit decisions	Correlators, Goertzel, thresholding

A recurring abstraction is that extraction is correct only relative to an invariant or canonical rule. In different papers, that rule is uniformity of a reservoir state, Zeckendorf nonadjacency, lexicographically higher preference, canonical IEEE-like decoding with software-defined bias, or uniqueness of an ANF after combining like terms modulo $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$0 (Mennucci, 2010, Borgohain, 2012, Yu et al., 2016, Agrawal et al., 2024).

2. Entropy-preserving extraction from random and pseudo-random sources

In RNG scaling, the problem is to transform independent $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$1-bit words $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$2 into independent uniform samples $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$3, where $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$4 may change on every draw. The naive map $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$5 is biased unless $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$6 divides $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$7. Writing $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$8 with $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$9, classes $$s = (x \gg (e + m))$$0 have $$s = (x \gg (e + m))$$1 preimages while $$s = (x \gg (e + m))$$2 have $$s = (x \gg (e + m))$$3. Exact single-word rejection therefore accepts only $$s = (x \gg (e + m))$$4, giving

$$s = (x \gg (e + m))$$5

The central refinement is bit recycling: maintain integers $$s = (x \gg (e + m))$$6 with invariant “$$s = (x \gg (e + m))$$7 is uniform on $$s = (x \gg (e + m))$$8.” After appending source words until $$s = (x \gg (e + m))$$9, compute $2$0 and $2$1. If $2$2, return $2$3, then set $2$4, $2$5; otherwise set $2$6, $2$7, and repeat. Unbiasedness is driven by quotient–remainder independence and by a process-splitting theorem. With a 64-bit reservoir kept in $2$8 and $2$9, the failure probability is at most $Y \in \{1,\dots,M\}$0; empirically, in $Y \in \{1,\dots,M\}$1 calls only three failures were observed, and measured entropy waste was $Y \in \{1,\dots,M\}$2 input bits over $Y \in \{1,\dots,M\}$3 input bits (Mennucci, 2010).

A different extraction setting appears in LFSR-based pseudo-random generators with output extraction. Three extractors are studied. The von Neumann extractor maps $Y \in \{1,\dots,M\}$4, $Y \in \{1,\dots,M\}$5, and discards $Y \in \{1,\dots,M\}$6 and $Y \in \{1,\dots,M\}$7, with expected output rate $Y \in \{1,\dots,M\}$8, hence $Y \in \{1,\dots,M\}$9 for $k$0. A 3-bit output-logic extractor consumes non-overlapping triplets and yields $k$1. A run extractor emits one bit when two consecutive LFSR bits differ, giving $k$2 for fair inputs; the paper states that the last one reaches $k$3. All three pass the FIPS 140-1 tests in the reported experiments, but they are described as less performant in terms of cryptographic strength than other generator logics (Nobach, 2024).

3. Bit planes, weighted decompositions, and image-domain extraction

In image-based steganography, bit extraction begins with bit-plane slicing. For an 8-bit grayscale pixel,

$k$4

and the $k$5-th plane is obtained by

$k$6

Classical LSB extraction is therefore a shift-and-mask scan over the stego image. Number-decomposition methods generalize the representation to weights $k$7, yielding “virtual” bit planes. In Fibonacci-weighted extraction, a greedy Zeckendorf decomposition enforces the rule that no two adjacent coefficients are both $k$8. In prime-number and natural-number decompositions, uniqueness is not intrinsic, so the paper uses the canonical rule that the “lexicographically higher number is given preference.” Capacity expands from 8 planes in classical LSB to 12 planes on 8-bit Lena for Fibonacci decomposition, 15 for prime numbers, and 23 for natural numbers (Borgohain, 2012).

Medical-image preprocessing uses bit-plane extraction in a different way. For 8-bit grayscale fracture X-rays, the paper constructs $k$9, $a_j$0, a fully denoised image, and a partially denoised image

$a_j$1

where only the two LSB planes are denoised with fastNLMeansDenoising. Full denoising gives the highest SNR and SSIM in the reported examples, but the partially denoised representation is emphasized for preserving higher-order structure while suppressing LSB noise. With DenseNet-based features and Random Forest, the partially denoised representation achieves $a_j$2 testing accuracy, while MSB4 reaches $a_j$3 in the same table (Paul et al., 21 Mar 2025).

In fully binarized neural networks, bit-plane encoding is used as an input-layer extraction technique. For an 8-bit input integer,

$a_j$4

and the front-end pipeline is Bit Rearrangement, Feature Extraction with depth-wise binary convolutions, Re-Weighting, and Fusion. The paper re-weights each bit-plane feature map by $a_j$5, implemented as a left shift, and then fuses over bit index:

$a_j$6

For first-layer complexity, the paper reports a ratio of $a_j$7 for $a_j$8, $a_j$9 for $y_i$0, and $y_i$1 for $y_i$2, with corresponding latency speedups of $y_i$3, $y_i$4, and $y_i$5. On CIFAR10, ReActNet-34 with $y_i$6 reaches $y_i$7 versus a $y_i$8 baseline (Vorabbi et al., 2023).

4. Symbolic bit recovery in algebraic and inverse problems

In GF$y_i$9 arithmetic verification, bit extraction means recovering a unique ANF for each output bit from a gate-level netlist with unknown irreducible polynomial $2$0. Gates are translated to polynomials over GF$2$1: XOR becomes addition modulo $2$2, AND becomes multiplication, OR becomes $2$3, and INV becomes $2$4. Starting from output wire $2$5, internal signals are recursively substituted by their gate polynomials; after each substitution, duplicate monomials cancel modulo $2$6. The result is a unique ANF for $2$7, and cancellations can occur only within the same output bit’s expression. Reconstruction of $2$8 then uses the degree-$2$9 signature

$N$0

For each output bit $N$1, the condition $N$2 is equivalent to $N$3. Because each bit is independent, extraction can run in $N$4 threads for an $N$5-bit multiplier. The reported experiments include Mastrovito multipliers with NIST polynomials up to $N$6, where the $N$7 case required $N$8 s and $N$9 GB, and flattened Montgomery designs up to $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$00, where the $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$01 case required $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$02 s and $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$03 GB (Yu et al., 2016).

In bit retrieval from periodic autocorrelation, extraction is an inverse phase-retrieval problem rather than a masking or decoding primitive. For a $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$04 sequence $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$05, periodic autocorrelation is

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$06

and in the Fourier domain

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$07

The task is to recover $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$08 from $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$09, possibly under noise. The most effective reported method is relaxed-reflect-reflect, with iteration

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$10

where $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$11 projects to the binary hypercube and $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$12 projects to the magnitude set. The paper reports an empirical optimum near $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$13, average-case growth $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$14 with $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$15, and hardest-instance growth with $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$16–$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$17. This usage of “bit extraction” is therefore explicitly an inverse-recovery problem constrained by autocorrelation and Fourier-magnitude structure (Elser, 2016).

5. Codecs, numeric formats, and storage-oriented extraction

In variable-byte integer codecs, extraction is centered on separating control bits from payload bits at SIMD width. VByte encodes each integer as bytes whose high bit is a continuation flag and whose low 7 bits carry payload, so

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$18

Masked VByte loads 16 bytes, applies pmovmskb to obtain a 16-bit MSB mask, uses the relevant 12-bit pattern to index a LUT, and applies PSHUFB to pack payload bytes into fixed positions. Logical masks strip the MSBs, and lane-wise shifts align $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$19-bit groups to weights $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$20. The implementation uses a two-level LUT with $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$21 primary entries and 170 shuffle masks, for a total of about $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$22 KB. On the reported benchmark, throughput ranges from about $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$23 million integers per second for the most compressible cases to about $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$24 million integers per second for the least compressible, versus about $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$25 down to about $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$26 for scalar VByte, i.e. a $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$27 to $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$28 speedup (Plaisance et al., 2015).

In arbitrary-bit-width quantization, eXmY generalizes IEEE-like floating point to $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$29 bits with exactly one sign bit, $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$30 exponent bits, and $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$31 mantissa bits. Extraction from a coded word $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$32 is

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$33

followed by IEEE-like decoding with a software-defined bias. For non-power-of-two widths, the codec uses a power-of-2 decomposition. For $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$34, the decomposition is $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$35: eight staged values are packed into one int32 lane holding the low 4 bits of all 8 elements, one int16 lane holding the next 2 bits, and one int8 lane holding the top 1 bit. Reconstruction for element $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$36 in an 8-pack is

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$37

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$38

The paper states that e3m1, a 5-bit format, gives about $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$39 compression from bfloat16, and that the technique has been deployed in production for almost 2 years (Agrawal et al., 2024).

A related storage-oriented extraction method represents exact 64-bit doubles in 32 bits by storing the sign, the 11-bit exponent, and the high-order 20 mantissa bits, then recovering the lower 32 mantissa bits by table lookup. If $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$40, the index is formed from low-order bits of $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$41 and, for some schemes, selected exponent bits. In Scheme B, for decimal data with 4 or fewer digits to the left and 2 or fewer to the right of the decimal point, $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$42, $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$43, so $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$44 and the table has $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$45 entries. In Scheme Z, for six decimal digits with the decimal point in any of seven positions, $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$46, $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$47, $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$48, so

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$49

requiring $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$50 direct entries, or $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$51 bytes, while an indirect form reduces the footprint to about $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$52 bytes (Neal, 2015).

6. Communication, bit-level permutation, and implementation constraints

In SDR demodulation, bit extraction is the final stage of a synchronization and detection chain. For coherent BFSK, matched filters or correlators compute

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$53

and a hard decision selects $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$54. Noncoherent variants use Goertzel bins or instantaneous-frequency slicing, and preamble detection can be done by correlating $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$55 with a known $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$56 preamble. The paper gives analytical BER formulas under AWGN:

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$57

for coherent BFSK and

$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$58

for noncoherent BFSK. The treatment is explicitly Python-centric, but the extraction logic is algorithmic rather than language-specific (Sanchez, 2024).

SMBBOT uses bit extraction as a reversible bit-level permutation over plaintext blocks. The input stream is partitioned into blocks of lengths $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$59, each block is written into an $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$60 square matrix in clockwise spiral order starting at the top-left cell, and the matrix is then decomposed into non-overlapping $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$61 sub-matrices scanned in row-major order. Bits are extracted column-wise from each $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$62 block to form ciphertext. Decryption places ciphertext bits column-wise into the $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$63 sub-matrices, reassembles the square matrix, and reads it in reverse of the spiral-placement process. The session key is the sequence of block-length and block-count pairs. In the reported experiments on 20 files, AES has the lowest encryption/decryption time, SMBBOT is faster than TDES, and TDES takes roughly three times the time of SMBBOT (Paul et al., 2013).

Across these implementations, the dominant correctness criteria are unbiasedness, canonicalization, reversibility, and hardware-aware efficiency. Bit recycling requires the invariant “$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$64 uniform on $$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$65” and exact acceptance-rejection; weighted image decompositions require Zeckendorf or lexicographic canonicalization; GF circuit extraction requires per-bit ANF uniqueness after modulo-$$P \approx \sum_{i=0}^{m-1} a_i w_i,\quad a_i\in\{0,1\},$$66 simplification; and eXmY lane packing requires byte-addressable reconstruction from standard storage types. The same literature also identifies recurring constraints: sender and receiver must use identical weight vectors, plane ordering, and canonicalization rules; branch-dependent accept/reject paths can matter in side-channel sensitive settings; and lookup-table or lane-decomposition designs are often chosen because they are amenable to SIMD or vector processing (Mennucci, 2010, Borgohain, 2012, Yu et al., 2016, Agrawal et al., 2024).